Cube Map to Equirectangular and Equirectangular to Cube Map OpenCV implementations with CUDA and OpenMP

Description

This repository allows you to transform equirectangular images to cube map images and vice versa with two types of implementations, both using C++ OpenCV one accelerated with OpenMP and the other accelerate with CUDA. It is a port from the Python version implemented here.

Requirements

• OpenCV 4.2.0<
• CUDA 10.1< (If using CUDA implementation)
• GCC 8
• CMake 2.8<

Build instructions

In order to be able to use this program, you need to build the project first. In order to do this, you need to go to the src/build and then modify the CMakeLists.txt file accordingly. You need to modify line 9 (find_package(OpenCV REQUIRED PATHS "/home/ema2159/.local/opencv/opencv-4.2.0/build")) changing the /home/ema2159/.local/opencv/opencv-4.2.0/build part with your own build of OpenCV. Then, you can remove the parts that you don’t need (I.E. the Include CUDA and CUDA implementation sections in case you only want OpenMP). Lastly, you need to execute the cmake . && make command inside the build directory. This will build all the executables inside the build directory which then you can utilize.

Usage instructions

In order to run any of the programs in question, you have to get into the build directory and execute the binaries from there as follows:

Equirectangular to cube map:

• OpenMP version:

./equrec2CubeMap ../../common/assets/sample.jpg ../../common/assets/sampleCubeMap2.jpg

• CUDA version:

./equrec2CubeMapCUDA ../../common/assets/sample.jpg ../../common/assets/sampleCubeMap2.jpg

Expected result:

Cube map to equirectangular:

• OpenMP version:

./cubeMap2Equrec ../../common/assets/sample- ../../common/assets/sample2.jpg

• CUDA version:

./cubeMap2EqurecCUDA ../../common/assets/sample- ../../common/assets/sample2.jpg

Expected result:

The default implementation only considers the case where the cube map is in six separate files. If you want to grab the cube map from a single image, go to the cubeMap2Equrec.cpp file (either in the OpenMP or the CUDA implementation according to your needs) and comment and uncomment the following section as follows:

   // Get six cube images inside a given directory
   // std::string imgs_path(argv[1]);
   // posY.upload(cv::imread(imgs_path + "posy.jpg"));
   // posX.upload(cv::imread(imgs_path + "posx.jpg"));
   // negY.upload(cv::imread(imgs_path + "negy.jpg"));
   // negX.upload(cv::imread(imgs_path + "negx.jpg"));
   // negZ.upload(cv::imread(imgs_path + "negz.jpg"));
   // posZ.upload(cv::imread(imgs_path + "posz.jpg"));

   // Extract images from cube map from a single file with the following format:
   //		+----+----+----+
   //		| Y+ | X+ | Y- |
   //		+----+----+----+
   //		| X- | Z- | Z+ |
   //		+----+----+----+
   cv::Mat h_img = cv::imread(argv[1]);
   posY.upload(h_img(cv::Rect(0, 0, h_img.cols / 3, h_img.rows / 2)));
   posX.upload(
	 h_img(cv::Rect(h_img.cols / 3, 0, h_img.cols / 3, h_img.rows / 2)));
   negY.upload(
	 h_img(cv::Rect(2 * h_img.cols / 3, 0, h_img.cols / 3, h_img.rows /
	 2)));
   negX.upload(
	 h_img(cv::Rect(0, h_img.rows / 2, h_img.cols / 3, h_img.rows / 2)));
   negZ.upload(h_img(cv::Rect(h_img.cols / 3, h_img.rows / 2, h_img.cols / 3,
				h_img.rows / 2)));
   posZ.upload(h_img(cv::Rect(2 * h_img.cols / 3, h_img.rows / 2, h_img.cols /
   3,
				h_img.rows / 2)));
   // Write individual extracted images
   cv::imwrite("posy.jpg",
		 h_img(cv::Rect(0, 0, h_img.cols / 3, h_img.rows / 2)));
   cv::imwrite("posx.jpg", h_img(cv::Rect(h_img.cols / 3, 0, h_img.cols / 3,
					    h_img.rows / 2)));
   cv::imwrite("negy.jpg", h_img(cv::Rect(2 * h_img.cols / 3, 0, h_img.cols /
   3,
					    h_img.rows / 2)));
   cv::imwrite("negx.jpg", h_img(cv::Rect(0, h_img.rows / 2, h_img.cols / 3,
					    h_img.rows / 2)));
   cv::imwrite("negz.jpg", h_img(cv::Rect(h_img.cols / 3, h_img.rows / 2,
					    h_img.cols / 3, h_img.rows / 2)));
   cv::imwrite("posz.jpg", h_img(cv::Rect(2 * h_img.cols / 3, h_img.rows / 2,
					    h_img.cols / 3, h_img.rows / 2)));